Rasyn
Making chemicals 100,000x faster
About Rasyn
We builds foundational AI models for chemistry. Almost nothing you buy is a single chemical. Products like datacenter coolant, GPU thermal paste, paint, and glue are formulations, which are mixtures of around ten ingredients each, and demand for them has grown quickly because of AI and EVs. Designing a new formulation is essentially a search problem, because there are more possible mixtures than anyone could ever test, and none of them can be predicted in advance. As a result, bringing a single formulation to market today takes three to five years of trial and error in a lab. Our models predict how a mixture will behave far faster than the standard methods used today, which lets us design new formulations in weeks rather than years, at a fraction of the cost. To prove this, we used them to invent three chemicals that never existed before, including a PFAS-free coolant for datacenters, and we have since synthesized and patented all three. We are already working with some of the largest manufacturers to design formulations for them, and we are building our own lab to generate the data that will train our next models.
Public traction evidence
Each signal links to the public source used for attribution.
- X
Today, we are launching IRIS, our open-source paper-finding agent.
Today, we are launching IRIS, our open-source paper-finding agent. Ask IRIS a literature question in plain English, with all of its conditions attached. It returns a ranked list of papers that were checked against every condition, one by one, with the sentences it relied on
- X
Today, we are launching Marigold.
Today, we are launching Marigold. Give it any chemistry task in plain language and it pulls the literature, reads your files, picks the tools, runs the calculations on its own machine, and hands back the numbers with the reasoning. Chemistry has countless software. One predicts
- X
Today, we are launching Analyst Agent, an AI that does analytical chemistry from the raw instrument file.
Today, we are launching Analyst Agent, an AI that does analytical chemistry from the raw instrument file. You give it the files an instrument produced and ask your question in plain language. It works out which measurements the question needs, runs them, and returns the answer
- X
Today, we are launching Larmor 2.8B, our NMR chemical shift prediction model.
Today, we are launching Larmor 2.8B, our NMR chemical shift prediction model. It is the most accurate published for proton shifts on carbon. NMR is how chemists find out what they actually made. You run a spectrum, you get a set of peak positions, and you check them against the
- X
Today, we are launching Marcus 670M, our reaction condition prediction model.
Today, we are launching Marcus 670M, our reaction condition prediction model. It is the best publicly reported model at this task. A reaction needs conditions to run: a catalyst, solvents, reagents. Choosing them is a separate problem from knowing what the reaction produces, and
- X
Today, we are launching Synthon 350M, our single-step retrosynthesis model.
Today, we are launching Synthon 350M, our single-step retrosynthesis model. Synthon is the strongest publicly reported model for this task when every model gets a single shot at the answer. Chemists make molecules by combining simpler ones. Retrosynthesis is that in reverse:
- X
Marigold now runs on your own machine.
Marigold now runs on your own machine. One command installs it: web app, agent core, Postgres, desktop icon. No GPU. If port 3000 is busy it picks a free one instead of failing. Bring your own model key and your provider is the only host your machine ever contacts. Chat
- X
Introducing MarigoldBench, a benchmark that measures whether an AI model can do the computational work that sits between a dataset and a decision in drug discovery.
Introducing MarigoldBench, a benchmark that measures whether an AI model can do the computational work that sits between a dataset and a decision in drug discovery. Each model gets 11 tools: ESMFold and Boltz-2 for structure prediction, RFdiffusion for backbone design,